Most pilots fail. Stopping them too early is the real mistake.
A company that tries several AI initiatives will see most of them fail to change much. That is not a sign the approach is wrong. It is how every new kind of investment behaves: most attempts do not work out, and a few work well enough to pay for all the others.
Expect a portfolio, not a single bet
Investors in new ventures have known this for decades. They do not expect each investment to succeed. They expect a small number to succeed by a wide margin, and they manage the whole set accordingly. Companies rebuilding processes around AI should think the same way: several bounded attempts, each small enough to lose, judged together.
The two ways to get it wrong
The first is to keep a failing pilot alive because stopping it feels like admitting defeat. Money and attention flow to the project that is not working, and the ones that might work are starved.
The second is worse and more common: stopping everything after the first failures, and concluding that AI does not work for the company. Costs come before results in any real change. A process run twice, the old way and the new, costs more for a while before the old way can be switched off. A board that reads that dip as proof of failure will stop exactly when the investment was about to pay.
How to run the set
Define the success criteria for each pilot before it starts, with a date. Stop the ones that miss clearly, quickly, and move the money to the ones that show progress. Write down what each failure taught, so the next attempt starts from there. And agree in advance, with the board, that a period of higher costs is part of the plan.
Failure is expected. Stopping too early is the mistake that costs the most.